Deep Ensemble Learning for Classification of Glaucoma from Smartphone Fundus Images

Glaucoma, a disease that damages the optic nerve, is a common cause of blindness in adults and requires an early diagnosis to delay the progress of the disease. In recent times, deep learning (DL) has been pivotal in classifying multiple retinal diseases including glaucoma from fundus images. In this study, we proposed an ensemble model (EM) based on multiple DL models for the classification of fundus images with glaucoma. Two distinct datasets were used in this experiment: LAG composed of good-quality and BrG composed of degraded quality images. We first tried to generalize multiple DL models by training on LAG, BrG, or combined LAG and BrG images independently. All deep learning (DL) models trained on the combined images maintained similar performance compared to DL models trained exclusively on LAG images when tested on LAG images. Moreover, the same models exhibited significantly improved performance when tested on BrG images compared to state-of-art DL models trained solely on BrG images. In addition, our proposed EM based on three DL models (VGG19, ResNetl8, and DenseNetl69) achieved better performance among any DL model, with an accuracy of 0.9697, sensitivity of 0.9701, and specificity of 0.9696 on the LAG dataset, and an accuracy of 0.9385, sensitivity of 0.9467, and specificity of 0.9305 on the BrG dataset. This demonstrates that the proposed EM exhibits generalization capabilities across diverse datasets, particularly in processing degraded image datasets. Furthermore, it shows promise as a valuable tool for detecting glaucoma from fundus images.